Single-shot S2S-ST in Spatial Transcriptomics

• The paper introduces a sample-specific paradigm that reconstructs dense spatial gene expression from very sparse data using a self-supervised sparser-to-sparse strategy.
• It leverages cascaded data consistency with hybrid attention networks and co-training with grayscale natural images to enhance imputation accuracy.
• Empirical results demonstrate significant improvements in MAE, PCC, and SSIM over existing methods, highlighting its potential in precise spatial transcriptomics reconstruction.

Single-shot Sparser-to-Sparse (S2S-ST) denotes a sample-specific learning paradigm in which a model is trained from a single sparsely sampled spatial transcriptomics specimen by making the available observation even sparser and learning to reconstruct the less sparse version; at inference, that learned mapping is used to impute a dense high-resolution expression map from sparse measurements. In its most explicit form, the term names the spatial transcriptomics framework introduced in 2025, but the underlying idea also has broader relevance to architectures that transform one sparse support into another sparse or denser task-relevant support in a single feed-forward or one-shot procedure (Fang et al., 22 Jul 2025, Gwak et al., 2020).

1. Definition, scope, and problem formulation

In the spatial transcriptomics setting, S2S-ST is designed for reconstructing high-resolution spatial gene expression maps when only a single sparsely sampled high-resolution ST specimen is available for training. The target is not conventional low-resolution-to-high-resolution enhancement from coarse spots. Rather, the task is to recover the full-resolution profile from a subset of already high-resolution sampling locations. That distinction is central to the formulation (Fang et al., 22 Jul 2025).

The phrase “single-shot” refers to the fact that the model is trained using only one ST sample from the target specimen. The model is therefore sample-specific: for each new tissue sample, training is repeated on that sample alone. The phrase “sparser-to-sparse” refers to the self-supervised strategy that manufactures supervision internally. The available sparse ST map is artificially made even sparser, and the model is trained to reconstruct the less sparse version from the more sparse version. The learned mapping is then reused at inference time to go from sparse observations to dense predictions. This is analogous in spirit to internal learning or self-supervised restoration in vision, but the formulation is expressed through explicit sampling masks and data-consistency constraints rather than paired dense labels.

For a target gene, the data are represented as image-like arrays over tissue space. The core patch notation is

$$X_h^{st}: \text{high-resolution ST patch of size } P \times P,$$

$$X_m^{st}: \text{medium-resolution sparse ST patch of size } P/S \times P/S,$$

$$X_l^{st}: \text{low-resolution, sparser ST patch, obtained by further downsampling } X_m^{st}.$$

The associated binary masks are $M_h^{st}$ and $M_m^{st}$. The imputation network $f_\theta$ is cascaded and produces stagewise outputs: $$\{\hat{X}^{st}_{m,1}, \hat{X}^{st}_{m,2}, \dots, \hat{X}^{st}_{m,K}\} = f_{\theta} (X_{l}^{st}, M_{m}^{st}),$$

$$\{\hat{X}^{st}_{h,1}, \hat{X}^{st}_{h,2}, \dots, \hat{X}^{st}_{h,K}\} = f_{\theta} (X_{m}^{st}, M_{h}^{st}).$$

A defining practical motivation is the cost and scarcity of ultra-high-resolution ST platforms. Xenium experiments are stated to cost roughly \$2,000–\$4,000 per sample excluding additional overhead, and can produce terabytes of data. A single-sample regime is therefore economically attractive but statistically difficult because it offers little training diversity, creates substantial overfitting risk, and provides no dense supervision at the missing sites one ultimately wants to recover.

2. Sample-specific training protocol and cross-domain co-learning

The end-to-end framework assumes two inputs: one sparse high-resolution ST whole-slide sample for a target gene, and a large collection of grayscale natural images from DIV2K used as an auxiliary co-training domain (Fang et al., 22 Jul 2025).

For ST data, each gene channel is processed separately. Expression values are log-normalized, the whole-slide map is cropped into $P \times P$ patches, and augmentation uses horizontal flips, vertical flips, and rotations by $$X_m^{st}: \text{medium-resolution sparse ST patch of size } P/S \times P/S,$$0. The high-resolution patch is then downsampled twice with stride $$X_m^{st}: \text{medium-resolution sparse ST patch of size } P/S \times P/S,$$1: $$X_m^{st}: \text{medium-resolution sparse ST patch of size } P/S \times P/S,$$2 The masks $$X_m^{st}: \text{medium-resolution sparse ST patch of size } P/S \times P/S,$$3 and $$X_m^{st}: \text{medium-resolution sparse ST patch of size } P/S \times P/S,$$4 record retained positions. In the reported implementation, $$X_m^{st}: \text{medium-resolution sparse ST patch of size } P/S \times P/S,$$5 and $$X_m^{st}: \text{medium-resolution sparse ST patch of size } P/S \times P/S,$$6, and the downsampling rule selects the top-left pixel of each $$X_m^{st}: \text{medium-resolution sparse ST patch of size } P/S \times P/S,$$7 grid.

For natural images, RGB DIV2K images are converted to grayscale and normalized to $$X_m^{st}: \text{medium-resolution sparse ST patch of size } P/S \times P/S,$$8. High-resolution patches $$X_m^{st}: \text{medium-resolution sparse ST patch of size } P/S \times P/S,$$9 are cropped and augmented similarly, except rotations can be arbitrary within $$X_l^{st}: \text{low-resolution, sparser ST patch, obtained by further downsampling } X_m^{st}.$$0. The same ST-derived masks are then imposed: $$X_l^{st}: \text{low-resolution, sparser ST patch, obtained by further downsampling } X_m^{st}.$$1 The mask sharing is significant because the natural-image branch is trained under the same sparse observation geometry as the ST branch.

Training is joint and uses a shared network $$X_l^{st}: \text{low-resolution, sparser ST patch, obtained by further downsampling } X_m^{st}.$$2 for both domains. For ST patches, the two tasks are: sparser-to-sparse, with input $$X_l^{st}: \text{low-resolution, sparser ST patch, obtained by further downsampling } X_m^{st}.$$3 and target $$X_l^{st}: \text{low-resolution, sparser ST patch, obtained by further downsampling } X_m^{st}.$$4; and sparse-to-dense via downsampling consistency, with input $$X_l^{st}: \text{low-resolution, sparser ST patch, obtained by further downsampling } X_m^{st}.$$5, prediction $$X_l^{st}: \text{low-resolution, sparser ST patch, obtained by further downsampling } X_m^{st}.$$6, and enforcement that downsampling the prediction with the known mask reproduces $$X_l^{st}: \text{low-resolution, sparser ST patch, obtained by further downsampling } X_m^{st}.$$7. For grayscale natural images, the corresponding tasks are fully supervised: $$X_l^{st}: \text{low-resolution, sparser ST patch, obtained by further downsampling } X_m^{st}.$$8 and $$X_l^{st}: \text{low-resolution, sparser ST patch, obtained by further downsampling } X_m^{st}.$$9. At inference, only the sparse ST sample is used. Dense high-resolution patches are predicted in a sliding-window manner over the whole slide, and overlapping predictions are fused by weighted averaging.

The paper’s modeling assumption is that natural images and ST expression maps share useful structural regularities, including local continuity, textures, edges, repeated patterns, and multi-scale spatial organization. This does not make the auxiliary domain biologically paired with ST. Rather, it provides dense supervision for a restoration task with matched masking geometry.

3. Cascaded Data Consistent Imputation Network

The architectural core of S2S-ST is the Cascaded Data Consistent Imputation Network (CDCIN), a cascaded model with $M_h^{st}$0 stages. Each stage contains a Data Consistency (DC) module and a Residual Dense Hybrid Attention Network (RDHAN). Before entering the cascade, the lower-resolution input is upsampled to high-resolution size, and each cascade predicts a residual refinement (Fang et al., 22 Jul 2025).

The refinement rule is

$M_h^{st}$1

$M_h^{st}$2

This makes CDCIN a residual iterative imputer rather than a single-stage direct regressor.

The DC module re-injects known measurements at every stage so that measured gene values are not overwritten. The paper writes

$M_h^{st}$3

with slightly inconsistent subscripts in the printed notation. The intended operation is clear: at known sampled positions, replace the current estimate with the observed value; at unknown positions, retain the current estimate. This data-consistency mechanism is central to the model’s biological interpretation, because measured spots are treated as fidelity constraints rather than merely noisy hints.

The default model uses $M_h^{st}$4 cascades. Ablation shows that performance improves significantly from 1 to 3 cascades, then saturates or slightly declines beyond that, making three cascades the reported accuracy-complexity trade-off.

Within each cascade, RDHAN is composed of multiple Residual Dense Hybrid Attention Blocks (RDHABs). Each RDHAB contains a Residual Dense Block (RDB) and a Hybrid Attention Block (HAB). The RDB uses dense connections and residual learning for feature reuse and gradient flow. The HAB combines channel attention through a Channel Attention-based Convolution Block (CAB) with spatial attention through Swin Transformer window-based self-attention. Its computation is

$M_h^{st}$5

$M_h^{st}$6

$M_h^{st}$7

The implementation uses CAB scale $M_h^{st}$8, window size 8, and 8 RDHAB blocks per RDHAN. This hybrid design is meant to capture both local fine-scale structure and longer-range spatial dependencies.

4. Self-supervised losses and the sparser-to-sparse learning mechanism

The central methodological innovation is the self-supervised construction of supervision from a single sparse ST sample (Fang et al., 22 Jul 2025).

Given a sparse whole-slide ST dataset, a high-resolution patch $M_h^{st}$9 is cropped, masked to yield a medium-resolution sparse patch $M_m^{st}$0, and then further downsampled to a sparser patch $M_m^{st}$1. The training pair is therefore generated internally: $M_m^{st}$2 This is the sparser-to-sparse task in the narrow sense of the term.

The ST self-supervised losses are

$M_m^{st}$3

$M_m^{st}$4

$M_m^{st}$5

The first term is the sparser-to-sparse self-supervised loss. The second is a sparse-to-dense self-supervised loss implemented through forward consistency: since the true dense target $M_m^{st}$6 is unknown, the model is not supervised directly against dense ground truth; instead, downsampling the dense prediction under the known sparse mask must reproduce the observed sparse patch $M_m^{st}$7. The ST loss excludes spots outside the tissue region.

The grayscale natural-image branch uses dense supervision with matched masking geometry: $M_m^{st}$8

$M_m^{st}$9

$f_\theta$0

The total objective is

$f_\theta$1

with $f_\theta$2.

The stage weights $f_\theta$3 emphasize later cascades. No adversarial loss, latent consistency penalty, or additional regularizer is reported. The regularization effect instead comes from the internal sparser-to-sparse task, the downsampling-consistency constraint, shared cross-domain training, and the architectural data-consistency mechanism.

5. Empirical performance, ablations, and reported implementation regime

Evaluation uses eight Xenium datasets from HEST-1K spanning five breast cancer samples, one prostate cancer sample, one healthy liver sample, and one diseased lymphoid tissue sample. The genes highlighted in experiments are ERBB2, CYP2A7, XBP1, and CCN1. The comparison includes BayesSpace, DIST, and TESLA. Metrics are Mean Absolute Error (MAE), Pearson Correlation Coefficient (PCC), and Structural Similarity Index (SSIM). Across all eight datasets, CDCIN/S2S-ST achieves the best MAE and SSIM, and usually the best PCC as well (Fang et al., 22 Jul 2025).

Illustrative results include the following. On TENX94 breast cancer for ERBB2, CDCIN reports MAE 0.3840, PCC 0.8539, and SSIM 0.7613, compared with DIST at 0.4470, 0.8467, and 0.7107. On TENX95 breast cancer for ERBB2, CDCIN reports 0.2801, 0.8025, and 0.7764. On TENX121 liver for CYP2A7, CDCIN reports 0.1889, 0.7243, and 0.8247. On TENX157 prostate cancer for CCN1, CDCIN reports 0.5264, 0.6224, and 0.8571. The qualitative description is that TESLA over-smooths, BayesSpace offers limited effective enhancement, DIST improves results but still misses sharp local variation, and CDCIN better preserves structural details and reduces local reconstruction error.

The ablation study isolates the roles of grayscale natural-image co-training and data consistency. Compared with plain RDHAN, adding GNI co-training improves performance consistently, with reported MAE reductions of roughly 4–7%. On TENX94, RDHAN yields MAE 0.4226, PCC 0.8216, and SSIM 0.7212; adding GNI co-training yields 0.4030, 0.8308, and 0.7380. Adding CDC alone often yields even larger gains. On TENX95, RDHAN yields 0.3164, 0.7581, and 0.7352, while adding CDC yields 0.2898, 0.7933, and 0.7721. Using both GNI and CDC gives the best overall results; on TENX94, the full CDCIN configuration yields 0.3840, 0.8539, and 0.7613.

The implementation details are unusually explicit. The reported configuration is: patch size $f_\theta$4; stride $f_\theta$5; three cascaded DC-RDHAN stages; 8 RDHAB blocks per RDHAN; RDB with 32 initial channels, growth rate 32, and 4 convolutional layers per block; optimizer Adam with default parameters; initial learning rate $f_\theta$6; training epochs 3000; hardware a single NVIDIA A6000 GPU; and training one gene at a time. The sample-specific training cost is about six hours of GPU time per new dataset.

6. Relation to adjacent sparse-support and sparse-to-sparser methods

Although the term S2S-ST is defined most explicitly for spatial transcriptomics imputation, related works show that the underlying idea of transforming one sparse support into another sparse or denser support in a single-shot pipeline appears in other domains.

A clear precursor is GSDN for 3D single-shot object detection. There, the input support consists of occupied surface voxels, while the desired detection support consists of candidate object-anchor locations that may lie in empty space. The method uses a hierarchical sparse tensor encoder, a generative sparse tensor decoder, transposed convolutions that expand support according to

$f_\theta$7

and learned pruning via sparsity prediction. The paper does not use the phrase S2S-ST, but it is an early example of a single-pass sparse-support transformation that remains sparse end-to-end rather than decoding to a dense voxel grid (Gwak et al., 2020).

Within spatial transcriptomics itself, ST-DAI is an adjacent but not strict match. It is sample-specific and single-shot in the sense of being calibrated solely on the specimen of interest, but its acquisition protocol assumes one fully sampled central section and sparsely sampled adjacent sections at roughly 25% spatial sampling density. Its goal is dense 3D reconstruction, not sparse-to-sparse reconstruction, and its pipeline relies on Cross-section Alignment, pseudo-map pretraining, Fast Multi-Domain Refinement, Parameter-Efficient Domain-Alignment Layers, a Confidence Score Generator, and a Data Consistency Operation that preserves measured adjacent values exactly. The paper is therefore best characterized as a one-full-plus-many-sparse to dense 3D reconstruction method rather than a pure S2S-ST formulation (Qian et al., 29 Jul 2025).

In model compression and efficient sequence modeling, the same conceptual pattern appears in different forms. SparseSSM introduces one-shot, training-free pruning for Mamba/selective SSMs and is naturally compatible with sparse-to-sparser pruning because saliency is computed over surviving parameters and then a new mask is applied in one shot. Its practical saliency for the time-shared transition parameter is

$f_\theta$8

which aggregates hidden-state statistics across time steps rather than using a purely local linear-layer score (Tuo et al., 11 Jun 2025).

SpenseGPT is also one-shot and post-training, but its target is hybrid rather than purely sparse: each weight matrix is partitioned into a dense region and a 2:4 sparse region so that execution can use existing dense and sparse GEMM libraries. It is therefore best viewed as a neighboring single-shot pruning framework whose endpoint is sparse+dense rather than sparse-to-sparser in a strict sense (Lee et al., 9 Jun 2026).

S2O is closer to the sparse-to-sparser archetype at inference time. It starts from block-wise sparse attention, constructs lightweight logical permutations $f_\theta$9 and $$\{\hat{X}^{st}_{m,1}, \hat{X}^{st}_{m,2}, \dots, \hat{X}^{st}_{m,K}\} = f_{\theta} (X_{l}^{st}, M_{m}^{st}),$$0, and traverses candidate blocks in estimated importance order. Early stopping is triggered when the online-softmax normalization gain satisfies

$$\{\hat{X}^{st}_{m,1}, \hat{X}^{st}_{m,2}, \dots, \hat{X}^{st}_{m,K}\} = f_{\theta} (X_{l}^{st}, M_{m}^{st}),$$1

so the remaining lower-priority blocks are skipped. In that sense, it turns an already sparse attention execution into a sparser realized computation within the same inference-time process (Zhang et al., 26 Feb 2026).

By contrast, “Sparser, Faster, Lighter Transformer LLMs” provides indirect rather than direct evidence for S2S-ST. It studies activation sparsity in Transformer FFNs induced during pretraining by ReLU and an $$\{\hat{X}^{st}_{m,1}, \hat{X}^{st}_{m,2}, \dots, \hat{X}^{st}_{m,K}\} = f_{\theta} (X_{l}^{st}, M_{m}^{st}),$$2 penalty, demonstrating that very high activation sparsity can be exploited with specialized kernels. This supports the broader feasibility of sparse computation, but it is not a one-shot sparse-to-sparse weight conversion method (Cetin et al., 24 Mar 2026).

7. Limitations, boundary conditions, and broader significance

The strongest current formulation of S2S-ST remains constrained by several explicit limitations (Fang et al., 22 Jul 2025).

First, retraining is sample-specific. Each new specimen requires a new training run, reported at roughly six hours on a single NVIDIA A6000 GPU. Second, the current implementation is single-gene-at-a-time rather than full-transcriptome reconstruction, so inter-gene dependencies are not modeled explicitly. Third, evaluation is mainly on technical reconstruction metrics—MAE, PCC, and SSIM—rather than downstream biological tasks such as differential expression, cell-type mapping, or biomarker discovery. Fourth, the transfer from grayscale natural images to ST maps is structural rather than semantic; the empirical ablations support it, but the domains are not biologically paired. Fifth, the framework is 2D only.

The broader literature suggests that the conceptual boundary of S2S-ST is still somewhat fluid. GSDN is sparse-to-denser-sparse rather than sparse-to-sparse at every intermediate step (Gwak et al., 2020). ST-DAI depends on one fully sampled reference section and therefore does not match an all-sparse single-shot regime (Qian et al., 29 Jul 2025). SparseSSM appears naturally compatible with sparse-to-sparser pruning, but does not directly benchmark already-sparse checkpoints (Tuo et al., 11 Jun 2025). SpenseGPT ends in a hybrid sparse+dense format rather than an all-sparse one (Lee et al., 9 Jun 2026). The Transformer activation-sparsity work is supportive at the systems level but indirect at the algorithmic level (Cetin et al., 24 Mar 2026). S2O is a strong inference-time sparse-to-sparser example, but it is specific to sparse attention and long-context prefill (Zhang et al., 26 Feb 2026).

Taken together, these works suggest two levels of meaning. In the narrow sense, S2S-ST refers to the histology-free, sample-specific framework for spatial transcriptomics imputation from a single sparse specimen, combining internal sparser-to-sparse supervision, cross-domain co-learning, and cascaded data consistency. In a broader methodological sense, it denotes a class of single-shot procedures that explicitly model the mismatch between observed sparse support and the target support required by the task, while preserving computational sparsity rather than reverting to dense reconstruction.